Information on cigarette smoking, an exposure of great epidemiological interest, is occasionally obtained from tumour registry data. We compared smoking histories from a hospital tumour registry with those from a questionnaire administered to 94 lung cancer patients. Reliability of tumour registry data was good for classifying individuals as ever-smokers and non-smokers (sensitivity 0.96, specificity 0.86). There was higher discrepancy in classifying smokers as current or former smokers. Current smokers had lower reliability for amount of smoking than former smokers, whereas both groups had high reliability for duration of smoking. These results suggest that tumour registry derived data on smoking must be used with caution.